SSE-YOLO: A Lightweight Model for Efficient Fire Detection in Real-Time Applications

The SSE-YOLO model, optimized for object detection, has exhibited exceptional effectiveness in fire detection tasks, particularly within environments where computational resources are limited, such as embedded devices and edge computing platforms. While the standard YOLOv8 model achieves commendable accuracy, its substantial computational load and memory requirements present obstacles for its adoption in embedded systems. In response to these limitations, this paper introduces the SSE-YOLO model, which integrates the SimAM module, a Slim-neck architecture, and a Shared Convolutional Detection head to enhance performance. The SimAM module is incorporated into the backbone network to improve feature extraction efficiency without increasing parameter counts. Additionally, the Slim-neck architecture is applied to streamline the model, reducing both computational burden and parameter complexity. Lastly, the Shared Convolutional Detection head is employed to further boost the model’s computational efficiency, thereby lowering resource usage during inference. Experimental evaluations on a fire detection dataset reveal that SSE-YOLO reduces model complexity and parameters by 45% and 31%, respectively, compared to YOLOv8n, while achieving improvements of 1.8% in mAP@0.5 and 3.8% in mAP@0.5:0.95. These findings demonstrate that SSE-YOLO delivers significant reductions in computational requirements while sustaining high detection performance, making it an ideal candidate for real-time fire detection on mobile and embedded systems.